Saturday, 5 September 2026
Artificial intelligence agents are beginning to move beyond experimental applications and into the everyday operations of large companies, but some of the organisations deploying them at scale are finding that the technology itself is only one part of the challenge. Data quality, employee trust, cost control and the ability to connect AI with measurable business outcomes are emerging as equally important factors determining whether projects move successfully from pilot programmes into production. That was the central message from an Ai4 2026 panel in Las Vegas bringing together executives working with AI across pharmaceuticals, healthcare, customer experience and digital businesses.
The discussion included Nitesh Soni of Sanofi, Anuj Maheshwari of CVS Health, Chris Han of Thinking AI and Kevin Lee of NiCE, with each describing how agentic systems are beginning to alter established business processes. The examples differed substantially, but a common pattern emerged. Successful projects were generally not being designed around the question of where a company could deploy an AI agent. They began instead with a business process that was slow, expensive or difficult to scale and then examined whether AI could improve the outcome.
At Sanofi, Soni described the development of a conversational data platform intended to give commercial teams faster access to information that previously required analysts to search through multiple dashboards and datasets. A brand manager or sales team might want to understand how a product is performing in a particular territory, where opportunities are emerging or where resources should be redirected. Traditionally, those questions can require analysts to combine information from several systems before producing an answer. The emerging model uses a network of specialised agents behind a conversational interface. According to Soni, the architecture currently involves roughly 10 to 12 agents, some performing general orchestration and others handling specific tasks. A user asks a question in ordinary language, an orchestration layer interprets the request and the appropriate agents retrieve and combine the necessary information.
The objective is not simply to generate an answer. The system can potentially suggest follow-up questions and recommend possible next actions, allowing employees without specialist analytics skills to interrogate enterprise information directly. This represents an important change from the traditional business-intelligence model, where organisations build dashboards and expect employees to understand where the relevant information sits and how it should be interpreted.
The CVS Health example takes a different approach. Maheshwari described what the company calls agentic twins, which are intended to simulate how different types of patients might respond to healthcare programmes before those programmes are introduced more widely. He said CVS Health has created hundreds of thousands of simulated agent profiles to test possible consumer responses while protecting patient privacy. The concept resembles a flight simulator for business decisions. Instead of launching a programme directly into the market and then discovering how patients react, an organisation can test different approaches in a virtual environment first.
Maheshwari connected the work to a longstanding healthcare challenge: patients who either do not collect prescribed medication or stop treatment earlier than intended. The purpose of the simulations is to understand the different barriers influencing those decisions and test whether alternative communications or programmes might improve outcomes before exposing real patients to the intervention. The significance extends beyond healthcare. Digital twins are already familiar in manufacturing, buildings and infrastructure, where companies create virtual representations of physical assets. Agentic AI introduces the possibility of creating behavioural simulations involving customers, employees or other groups, allowing companies to test decisions before committing resources in the physical marketplace.
For businesses, this could change the economics of experimentation. Marketing programmes, customer experiences, pricing strategies and operational changes could potentially be tested against large synthetic populations before companies launch them. The technology is still developing, and simulated behaviour cannot automatically be assumed to reproduce real human decisions accurately, but the approach demonstrates how agentic AI could move beyond task automation into decision modelling.
NiCE’s Kevin Lee approached the issue from the customer-experience perspective. His argument was that companies frequently begin with the technology rather than examining what customers actually need. A business may decide that a generative-AI model should solve a particular problem even when a conventional rules-based system or established automation could perform the task more reliably and at lower cost. That distinction becomes increasingly important as companies begin paying for AI usage at scale. A simple request such as checking the status of an order may not require an advanced generative model, while more complicated conversations involving several intentions, changing context or unstructured information may justify a more capable system.
The emerging enterprise architecture is therefore unlikely to involve sending every problem to the largest available AI model. Companies are instead beginning to route different tasks to different technologies depending on complexity, cost and risk. Some processes may use conventional software, others smaller AI models, while difficult reasoning tasks are sent to more powerful systems. An orchestration layer determines which approach is appropriate. This could become one of the most important economic questions surrounding enterprise AI because individual interactions with models can appear inexpensive, but the cost becomes considerably more significant when millions of customers, employees or automated agents are generating requests continuously.
The objective therefore changes from maximising the intelligence used for every task to finding the least expensive architecture capable of producing the required outcome reliably. The panel repeatedly returned to one factor underlying all of these systems: data. Soni argued that organisations unable to establish reliable, governed information foundations will struggle to scale agentic AI regardless of how capable the underlying models become. A prototype can sometimes operate on carefully prepared datasets, but enterprise deployment has to work across information created by different departments, systems and historical processes.
Large companies frequently have several versions of the same information stored in different places. Documents may be outdated, databases may use inconsistent definitions and responsibility for maintaining particular information may have disappeared as employees changed roles. AI does not automatically solve those problems. Connecting a powerful model to five conflicting sources does not necessarily produce a better answer. It may simply allow the system to choose between five different versions of reality. This is why the less visible work of cataloguing, governing and maintaining enterprise data is becoming increasingly important to AI adoption.
Sanofi’s approach, as described by Soni, separates information into different layers. Raw information from source systems is first organised before being aligned with business processes and eventually transformed into datasets suitable for analytics and AI applications. AI itself can also help perform some of that preparation, creating a cycle in which better data enables better AI while AI assists companies in improving their data.
Trust becomes the next challenge. An AI system can produce a highly accurate recommendation, but the business value remains limited if employees do not understand the result or refuse to act on it. This makes explainability and transparency important not only for regulatory reasons but for adoption. Employees need to understand where information came from, why a recommendation was produced and whether they remain responsible for the final decision. This is particularly important in healthcare and pharmaceuticals, where incorrect recommendations can have consequences far beyond lost productivity.
Systems therefore need observability, testing and traceability designed into them from the beginning rather than added after deployment. Maheshwari emphasised the need to distinguish between what AI is technically capable of doing and what an organisation should allow it to do. Human intelligence remains essential where judgement, safety or accountability are involved. The principle applies beyond healthcare because companies deploying AI agents increasingly need to decide which actions can be automated, which require human approval and which should not be delegated to AI at all.
Change management is another major part of the equation. Soni argued that business teams need to become involved from the beginning rather than being presented with a finished AI product and expected to adopt it. Continuous feedback from users can improve the technology while also creating internal advocates who understand why the system was built and how it should be used. This suggests that scaling AI may ultimately be as much an organisational problem as a technical one. Companies can purchase access to increasingly capable models relatively easily, but changing how thousands of employees work, how decisions are made and how responsibility is distributed across the organisation is considerably more difficult.
The measurement of success is also becoming more sophisticated. Early AI programmes frequently concentrated on technical accuracy or the number of tasks automated. The panel argued that companies increasingly need to connect those measurements with operational and financial outcomes. One useful metric is time to insight. Maheshwari described analytical work that could previously require several people working for days being reduced dramatically through AI-assisted processes. The relevant question, however, is not simply whether the AI completed the task faster but whether it maintained the required standards of accuracy, safety and reliability.
In pharmaceuticals, time can have even greater significance. Soni argued that reducing the time required across processes ranging from coding and commercial analysis to bringing therapies to patients can produce meaningful value when improvements are multiplied across a global organisation. Customer-experience businesses use different measurements, including satisfaction, successful resolution and the cost of individual interactions, while digital businesses may concentrate more heavily on customer retention and engagement. The common principle is that an AI project should ultimately be measured against the business problem it was supposed to solve.
This addresses one of the central difficulties surrounding corporate AI investment. The rapid development of generative and agentic AI has encouraged companies to experiment widely, sometimes without defining the financial or operational result expected from the project. As enterprise spending increases, that tolerance is likely to decline. Executives will increasingly ask how much a system costs to operate, which processes it improves, how much employee time it saves and whether those improvements ultimately influence revenue, margins, customer retention or another strategic objective.
Cost management is therefore becoming part of AI architecture itself. The panel described an emerging combination of buying, renting and internally developing AI capabilities. Companies may purchase common platforms for widely used functions, access specialist models or tools where capabilities change rapidly and develop their own systems where proprietary knowledge provides a competitive advantage. Maintaining flexibility could become important because the AI market is evolving too quickly for companies to assume that today’s leading model or architecture will remain dominant.
The same principle applies inside individual workflows. Not every task requires generative AI, and not every generative task requires the most sophisticated model available. Matching computational cost to the complexity and economic value of the task could become a basic discipline of enterprise AI management. The wider implication is that AI agents are beginning to change business architecture rather than merely automate individual jobs.
Companies historically organised workflows around humans moving information between software applications. An employee opened a dashboard, interpreted the information, transferred it into another system, made a decision and then initiated the next action. Agentic systems can increasingly connect those stages. One agent can retrieve information, another analyse it, another recommend an action and another execute an approved step, while humans increasingly supervise the workflow, resolve exceptions and make higher-value judgements.
That does not mean every company is close to autonomous operations. The panel repeatedly acknowledged that enterprise deployment remains difficult and that many organisations are still learning how to move projects beyond controlled pilots. But the direction is becoming clearer. The most successful enterprise AI strategies may not be those with the largest number of agents or the most advanced models. They may instead be those with the strongest information foundations, clearest business objectives and best understanding of where machines should operate independently and where humans should remain involved.
For companies investing heavily in AI, this represents an important shift in priorities. The question is moving away from how many AI tools an organisation has deployed towards whether those systems are improving the economics and effectiveness of the business. AI agents can make information easier to access, simulate possible decisions, automate workflows and reduce the time required to analyse complex situations, but none of those capabilities automatically creates value.
The underlying data still needs to be reliable, employees still need to trust the system, costs still need to be controlled and outcomes still need to be measured. As agentic AI moves deeper into large organisations, those fundamentals may ultimately determine which companies turn the technology into a genuine operating advantage and which simply build another layer of expensive software.
Source: CIJ.World Research & Analysis Team
